3M-Net: Automatic Modulation Recognition Based on Multiscale Mobile Inverted Bottleneck Convolution and Manhattan Self-Attention Network

Abstract

The increasingly complex modern communication environment poses challenges for automatic modulation recognition (AMR) techniques. In AMR tasks, in order to more comprehensively capture signal features and improve recognition performance, we propose a model named Multiscale Mobile Inverted Bottleneck Convolution and Manhattan Self-Attention Network (3M-Net). In this 3M-Net, the MSMB block is designed to extract multiscale local features of the signals, and the MMG block is designed to enhance global information modeling of the model. Then, a hierarchical backbone that contains the two blocks is designed to extract multilevel features. Extensive experiments on the RML2016.10a and RML2018.01a datasets demonstrate that the 3M-Net model achieves superior recognition performance.